Memory network with hierarchical multi-head attention for aspect-based sentiment analysis

Aspect-based sentiment analysis is a challenging subtask of sentiment analysis, which aims to identify the sentiment polarities of the given aspect terms in sentences. Previous studies have demonstrated the remarkable progress achieved by memory networks. However, current memory-network-based models cannot fully exploit long-term semantic relationships to the given aspect terms in sentences, which may lead to the loss of aspect information. In this paper, we propose a novel memory network with hierarchical multi-head attention (MNHMA) for aspect-based sentiment analysis. First, we introduce a semantic information extraction strategy based on the rotational unit of memory to acquire long-term semantic information in context and build memory for the memory network. Second, we propose a hierarchical multi-head attention mechanism to preserve aspect information and enable MNHMA to focus on the critical context words to the given aspect terms in sentences. Third, we employ a fully connected layer in each attention layer of the hierarchical multi-head attention layer to simulate the nonlinear transformation of sentiments, thereby acquiring a comprehensive context representation for aspect-level sentiment classification. Experimental results on three commonly used benchmark datasets demonstrate that our MNHMA model outperforms other state-of-the-art models for aspect-based sentiment analysis.